Blast furnace molten iron silicon content soft measurement method based on Copula function and space-time convolutional network

Abstract

The invention provides a blast furnace molten iron silicon content soft measurement method based on a Copula function and a space-time convolution network. In combination with an R-vine copula function and a graph neural network (GCN), proposing an inter-process-variable feature extraction framework depending on blast furnace data, obtaining an inter-variable relationship, and further realizing extraction of inter-process-variable features of the blast furnace; the method comprises the following steps of: performing time sequence feature extraction on a blast furnace process variable sequence by using a time convolutional neural network (TCN) with a gating mechanism according to a process variable time sequence feature extraction framework of blast furnace data; and splicing the extracted inter-process variable characteristics, the extracted time sequence characteristics and the extracted original process variable sequence, and realizing soft measurement of the silicon content of the molten iron of the blast furnace by adopting a regression device formed by multiple linear layers. According to the method, important details of original data can be prevented from being lost when GCN and TCN are used for feature extraction, and meanwhile, the accuracy and robustness of a molten iron soft measurement model can be enhanced.

Description

Technical Field

[0001] The invention relates to the field of blast furnace ironmaking molten iron quality prediction, and in particular to a soft measurement method for silicon content in blast furnace molten iron based on Copula function and spatiotemporal convolutional network. Background Art

[0002] In the blast furnace ironmaking process, there are many variables that characterize the process state that are constantly changing, including state variables, quality variables, and operating variables. The prediction of quality variables is crucial for process optimization, fault detection, diagnosis and prevention, and operational decision-making. Among these quality variables, silicon content, as a key indicator reflecting the internal chemical heat of molten iron, is one of the most important variables. With the continuous improvement of the industrial modernization of the ironmaking process, the measurement problems of many state variables have been basically solved, but due to the harsh industrial environment, some variables that are critical to product quality cannot be measured directly or the measurement cost is too high. Therefore, it is even more important to perform soft measurement on variables that are difficult to measure directly.

[0003] The blast furnace ironmaking process is a metallurgical process that converts iron ore into molten iron. This process is a continuous production method, that is, the raw materials interact from top to bottom, while the gas moves in the opposite direction from bottom to top. The blast furnace ironmaking process involves the proportional mixing and grinding of solid raw materials such as iron ore, limestone and coke. Then, these materials are thrown into the upper part of the blast furnace. Preheated air and coal powder are injected into the blast furnace and burned to produce high-temperature and high-pressure reducing gases. As the gases rise, they heat the charge in the furnace, resulting in a series of complex physical and chemical transformations. These reactions reduce the iron oxide in the iron ore to molten iron, which can be collected through the iron mouth. In addition, limestone and other substances are injected into the blast furnace hearth to react with impurities in the charge and promote the separation of slag. Since blast furnace exhaust and flue gases contain a lot of thermal energy, they are usually collected to generate electricity or heat the charge. Faced with such a complex blast furnace ironmaking process, its dynamics, complex nonlinearity, numerous variables, noise and outliers make the soft measurement task of molten iron quality indicators face great challenges.

[0004] In the blast furnace ironmaking process, if the measurement of silicon content in molten iron mainly relies on manual testing, there will be obvious disadvantages. First of all, the testing operation often has obvious lag, which has an adverse impact on the subsequent blast furnace operation optimization and product quality management. However, with the continuous improvement of the level of industrial informatization, data-driven quality indicator soft measurement technology has become a key technology to improve quality and operational efficiency. Summary of the invention

[0005] In order to overcome the shortcomings of the existing methods, the purpose of the present invention is to provide a soft measurement method for silicon content in molten iron of a blast furnace based on Copula function and spatiotemporal convolutional network.

[0006] The technical solution of the present invention to achieve its technical purpose is as follows:

[0007] A soft measurement method for silicon content in molten iron of blast furnace based on Copula function and spatiotemporal convolutional network is proposed. Combining R-vine copula function and graph neural network (GCN), a feature extraction framework between process variables dependent on blast furnace data is proposed to obtain the relationship between variables, and then realize the extraction of features between blast furnace process variables; a time series feature extraction framework for process variables dependent on blast furnace data is used to extract the time series features of blast furnace process variable sequence using a time convolutional neural network (TCN) with a gating mechanism; the extracted process variable features, time series features and original process variable sequence are spliced, and a regressor composed of multiple linear layers is used to realize the soft measurement of silicon content in molten iron of blast furnace.

[0008] The inter-variable feature extraction framework that relies on blast furnace data is as follows: the rattan copula fitting process includes constructing a bivariate copula using the Bayesian formula, calculating a conditional probability formula based on the assumed bivariate copula, and selecting and optimizing the rattan copula model based on maximum likelihood; constructing an adjacency matrix between variables based on the tree structure matrix generated during the rattan copula fitting process, and then combining a graph convolutional neural network to perform variable feature extraction.

[0009] The process variable time series feature extraction framework relying on blast furnace data uses two parallel connected TCN modules and different gating units, and uses sigmoid and tanh functions as activation functions to achieve reasonable time feature extraction.

[0010] The present invention has the beneficial effects:

[0011] 1. A feature extraction framework between variables is proposed by combining the R-vine copula function and graph neural network (GCN) to obtain a more comprehensive relationship between variables and achieve efficient feature extraction, aiming to deal with the nonlinear and dynamic characteristics of the blast furnace ironmaking process.

[0012] 2. Use a temporal convolutional neural network (TCN) with a gating mechanism to extract time series features of blast furnace process variable sequences and enhance the ability to capture long time scale dependent features.

[0013] 3. Splice the extracted inter-variable features, time series features and original process variable sequences to ensure the integrity of important information extracted from the features. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 Shown is a schematic diagram of a framework of a soft measurement method for silicon content in molten iron in a blast furnace based on a Copula function and a spatiotemporal convolutional network according to the present invention.

[0015] Figure 2The experimental results of VC-TGCN soft measurement of silicon content in molten iron.

[0016] Figure 3 The figure shows the test effect of the soft measurement of the method of the present invention on the molten iron silicon content data set. DETAILED DESCRIPTION

[0017] In order to more clearly describe the technical content of the present invention, the specific implementation method is further explained below in conjunction with the accompanying drawings.

[0018] like Figure 1 As shown, it is a schematic diagram of the framework of a soft measurement method for silicon content in molten iron of a blast furnace based on Copula function and spatiotemporal convolutional network of the present invention, and the steps are as follows:

[0019] Combining the R-vine copula function and the graph neural network (GCN), a variable feature extraction framework that depends on blast furnace data is proposed to obtain a more comprehensive relationship between variables, thereby realizing the efficient extraction of features between blast furnace process variables. In addition to extracting features between blast furnace process variables, a process variable time series feature extraction framework that depends on blast furnace data is used. This framework uses a temporal convolutional neural network (TCN) with a gating mechanism to extract time series features of blast furnace process variable sequences. By splicing the extracted variable features, time series features, and original process variable sequences, the integrity of important feature extraction information is ensured. Then, a regressor composed of multiple linear layers is used to realize the soft measurement of silicon content in blast furnace hot metal.

[0020] The inter-variable feature extraction framework that relies on blast furnace data is:

[0021] Copula function is widely used in multivariate joint distribution modeling. It can connect the cumulative probability distribution (CDF) of each variable to be analyzed into their joint distribution. The traditional two-variable copula function needs to fix the copula function family used in advance when modeling the joint probability distribution of multivariate, which cannot effectively model the blast furnace process variables with strong dynamics. The vine copula decomposes the multivariate copula into a series of two-variable copulas by establishing multiple tree structures. The actual vine copula fitting process includes constructing a two-variable copula using the Bayesian formula, calculating the conditional probability formula based on the assumed two-variable copula, and selecting and optimizing the vine copula model based on maximum likelihood.

[0022] Let x1,x2,…,x m There are m blast furnace process variables, then the joint probability distribution f(x) of these m process variables can be decomposed according to the following formula:

[0023]

[0024] After the above decomposition, the key item of copula function fitting becomes the conditional probability of a variable under the conditions of other variables. This conditional probability can be solved according to the following formula:

[0025]

[0026] When variable number i=t-1, the formula

[0027] c m-i,m∣1:m-i-1 (F(x m-i |x1,…,x m-i-1 ),F(x m |x1,…,x m-i-1 ));

[0028] It can be simply expressed as c 1,t (F(x1),F(x t )), and the formula describes the steps of using Rvine copula to decompose the joint probability distribution. After decomposition, the type of copula function can be selected according to the fitting effect. In order to fit the actual blast furnace process variable distribution characteristics, the present invention limits the copula function family to five copula function families: Gaussian, Student, Clayton, Gumbel and Frank. The specific data expression is expressed in the following table:

[0029]

[0030]

[0031] Since the bivariate copula is fitted by processing paired variables, the m process variables are decomposed into The generalized modeling formula using R copula can be expressed as follows:

[0032]

[0033] in Refers to removing x i and x j Based on the above formula, the CDF conditional distribution mentioned above can be solved step by step in an iterative manner according to the order of the tree structure. After calculating the conditional distribution of CDF, the overall framework has been established. The next thing to do is to estimate and optimize The parameters of the copula and its position in the tree structure. The estimation and optimization process uses the maximum likelihood estimation method (MLE). The traditional optimization formula is to parameterize the marginal distribution and optimize the marginal distribution parameters together with the parameters required for the copula fitting process. In order to simplify the process, the optimization of the marginal distribution and the optimization of the copula fitting process parameters are performed separately. The marginal distribution of the variables is modeled using kernel density estimation (KDE), and then the generalized copula parameter optimization formula is obtained based on the maximum likelihood method and the Akaike information criterion (AIC), which is as follows

[0034]

[0035] in represents the parameters of all copulas, Represents the position in the tree structure formed by the copula function, which is called the tree order. Finally, after the fitting process, we will get m-1 trees, each tree represents a set of dependencies, such as the edges of the first tree represent the direct connection between variables, the edges of the second tree represent the direct connection between variables, and the edges of the third tree represent the direct connection between variables. The edges of the tree represent the direct connection between variables, and the edges of the second tree represent conditional dependencies. The edges of the second tree represent the conditional dependencies between variables given the first tree, and so on. The conditional dependencies between variables, and the subsequent trees are deduced accordingly. And so on. Starting from the first tree, the edges of each tree will become nodes of the next tree. Starting from the first tree, the edges of each tree will become nodes of the next tree. Therefore, the roles of nodes and edges are constantly changing to build higher-level conditional dependencies. In order to briefly show the process of building a tree structure, Figure 2 This is the result of using R vine to build a tree structure for five blast furnace process variable data. The number of the tree structure represents the order in which the copula fits different variables. The bottom tree structure is the result of combining 1, tree 2, tree 3 and tree 4, that is, the vine structure. Each row in the figure above represents a pair of copulas connecting two variables. After determining the overall structure, the categories and parameters are also determined. In other words, using the tree structure, we can create the adjacency matrix required for GCN in the network below.

[0036] The framework for extracting time series features of process variables that rely on blast furnace data is:

[0037] Compared with recurrent neural network (RNN) and long short-term memory neural network (LSTM), time convolution neural network (TCN) has the unique advantage of parallel processing, so the time series feature extraction process of blast furnace molten iron silicon content related information is carried out using TCN network. At the same time, in order to further enhance the ability of TCN to capture long-time scale dependent features, the very important gating mechanism in LSTM and gated recurrent unit (GRU) is introduced into the time convolution network, such as forget gate, input gate and output gate. It is found through experiments that the time convolution network does not need input gate and forget gate, so an output gating unit is added at the end of the time series convolution neural network. Experiments show that this output gating unit similar to LSTM can more effectively analyze the long-range spatiotemporal dependency and extract spatiotemporal features. Therefore, when designing the TCN network, the present invention uses two parallel connected TCN modules and different gating units, and uses sigmoid and tanh functions as activation functions to achieve reasonable time feature extraction.

[0038] The specific process of the method provided by the present invention is:

[0039] Step 1: Offline training;

[0040] Step 1.1: In the soft measurement of silicon content in molten iron based on VC-TGCN, the mean square error (MSE) is used as the model training loss type, which can be given by the following formula:

[0041]

[0042] Among them, y i and Represents the soft measurement value and true value of the silicon content in molten iron output by the model at time i. The model is updated according to the MSE loss.

[0043] Step 2: Real-time prediction;

[0044] First, real-time data is collected and standardized, and the real-time soft measurement value of molten iron silicon content is output through the trained VC-TGCN model.

[0045] Substituting the present invention into actual industrial data for verification, we take a certain ironmaking plant with a volume of 2650m 3 The data sample of silicon content in molten iron of No. 2 blast furnace in 2020 contains 28 parameters. 600 normal samples are used for training, and 200 points are used as test sets for testing. Figure 3 The figure shows the test effect of the soft measurement of the method of the present invention on the molten iron silicon content data set. It can be seen from the soft measurement results that the model effect is good.

[0046] The technical features of the above-described embodiments can be further combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0047] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the invention. It should be pointed out that for ordinary technicians in this field, several modifications and improvements can be made without departing from the concept of the present invention, all of which belong to the protection scope of the present invention. The protection scope of the present invention is given by the attached claims and any equivalent technical solutions thereof.

Claims

1. A soft measurement method for silicon content in molten iron of a blast furnace based on Copula function and spatiotemporal convolutional network, characterized in that: Combining the R-vine copula function and graph neural network (GCN), a feature extraction framework between process variables that depends on blast furnace data is proposed to obtain the relationship between variables, and then realize the extraction of features between blast furnace process variables; a time series feature extraction framework for process variables that depends on blast furnace data uses a temporal convolutional neural network (TCN) with a gating mechanism to extract the time series features of the blast furnace process variable sequence; the extracted process variable features, time series features and original process variable sequence are spliced, and a regressor composed of multiple linear layers is used to realize soft measurement of silicon content in blast furnace molten iron.

2. The method according to claim 1, characterized in that The inter-variable feature extraction framework that relies on blast furnace data is: The vine copula fitting process includes constructing a bivariate copula using the Bayesian formula, calculating the conditional probability formula based on the assumed bivariate copula, and selecting and optimizing the vine copula model based on maximum likelihood; constructing the adjacency matrix between variables based on the tree structure matrix generated during the vine copula fitting process, and then combining the graph convolutional neural network to extract variable features.

3. The method according to claim 1, characterized in that The process variable time series feature extraction framework relying on blast furnace data uses two parallel connected TCN modules and different gating units, and uses sigmoid and tanh functions as activation functions to achieve reasonable time feature extraction.

Images